The IASLAB-RGBD Fallen Person Dataset consists of several RGB-D frame sequences containing 15 different people. It has been acquired in two different laboratory environments, the Lab A and Lab B. It can be divided into two parts: the former acquired from 3 static Kinect One V2 placed on 3 different pedestals ; the latter from a Kinect One V2 mounted on our healthcare robot prototype, see "An Open Source Robotic Platform for Ambient Assisted Living" by M. Carraro, M. Antonello et al in AIRO at AI*IA 2015.
Both parts are briefly described in the following. They contain the training/test splits of our approach to detect fallen people.
- Folder "raw": 360 RGB frames and point clouds with the camera calibrations;
- Folder "segmented_fallen_people": point clouds of the fallen people. They have been manually segmented;
- Folder "training_with_cad_room_and_nyudv2": random selected positives (70%), 24 point clouds from the Lab A and 31 point clouds from the NYU Depth Dataset V2 by Silberman et al;
- Folder "test_with_lab_room": random selected positives (30%) and 32 point clouds from the Lab B.
- Folder "training": 4 ROS bags with 15932 RGB-D frames in total acquired during 4 robot patrollings of the Lab A;
- Folder "test": 4 ROS bags with 9391 RGB-D frames in total acquired during 4 robot patrollings of the Lab B. This room is more similar to an apartament: spaces are smaller, it is cluttered and contains a sofa;
- Folder "maps": 2D maps of the two environments and ground truth positions of the person centroids in the maps.
In the figures below, some RGB samples from both environments are reported: